Abstract:

To predict the possibility of traffic accidents in advance， a traffic accident risk anticipation model gated recurrent unit spatial-temporal transformer（GST） was established based on the combination of vision transformer（ViT）， gated recurrent unit （GRU）， and MLP-Mixer. By modeling spatial-temporal relational learning through ViT， the frame features of predicted targets were enhanced to improve their distinguishability. On this basis， GRU was used to extract temporal relational， and then GRU and MLP-Mixer were combined to enhance the hidden layer frame features， establishing and optimizing spatiotemporal relational， the confidence score of traffic accidents for each time step was predicted based on the corresponding feature frames to predict the probability of future accidents and effectively distinguish between dangerous driving and accident driving behavior. Finally， the proposed model was validated on the public datasets DAD and A3D， and the results showed that the recognition accuracy of the proposed model was superior to other advanced algorithms. The AP on the two datasets reached 59.9% and 94.6%， respectively， demonstrating good predictive performance and generalization ability. In the DAD dataset， the algorithm proposed was compared to the DSTA model. With similar AP， the proposed algorithm can advance the prediction time of accidents by 2.38 seconds， an increase of about 13%. This indicates a significant advantage and provides assistance for road hazard warnings and safe driving.

Key words: intelligent transportation, accident anticipation, event discriminator, gated recurrent unit, spatial-temporal relational